deepseek_v2.py 48.4 KB
Newer Older
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3

wangding zeng's avatar
wangding zeng committed
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
# Adapted from
# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/llama/modeling_llama.py
# Copyright 2023 The vLLM team.
# Copyright 2023 DeepSeek-AI and the HuggingFace Inc. team. All rights reserved.
#
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
# and OPT implementations in this library. It has been modified from its
# original forms to accommodate minor architectural differences compared
# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
25
"""Inference-only DeepseekV2/DeepseekV3 model."""
王敏's avatar
王敏 committed
26
27
import os
import re
28
import vllm.envs as envs
zhuwenwen's avatar
zhuwenwen committed
29

30
31
import typing
from collections.abc import Callable, Iterable
32
from typing import Any, Optional, Union
wangding zeng's avatar
wangding zeng committed
33
34
35
36
37

import torch
from torch import nn
from transformers import PretrainedConfig

38
from vllm.attention import Attention
39
from vllm.compilation.decorators import support_torch_compile
40
41
from vllm.config import (CacheConfig, ModelConfig, VllmConfig,
                         get_current_vllm_config)
王敏's avatar
王敏 committed
42
from vllm.distributed import (get_ep_group, get_pp_group, get_dp_group,
43
                              get_tensor_model_parallel_world_size)
wangding zeng's avatar
wangding zeng committed
44
from vllm.model_executor.layers.activation import SiluAndMul
45
from vllm.model_executor.layers.fused_moe import FusedMoE
wangding zeng's avatar
wangding zeng committed
46
47
48
49
50
51
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
                                               MergedColumnParallelLinear,
                                               ReplicatedLinear,
                                               RowParallelLinear)
from vllm.model_executor.layers.logits_processor import LogitsProcessor
52
from vllm.model_executor.layers.quantization import QuantizationConfig
wangding zeng's avatar
wangding zeng committed
53
54
55
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.vocab_parallel_embedding import (
    ParallelLMHead, VocabParallelEmbedding)
56
57
from vllm.model_executor.model_loader.weight_utils import (
    default_weight_loader, maybe_remap_kv_scale_name)
wangding zeng's avatar
wangding zeng committed
58
from vllm.model_executor.sampling_metadata import SamplingMetadata
59
from vllm.sequence import IntermediateTensors
wangding zeng's avatar
wangding zeng committed
60

61
from .interfaces import MixtureOfExperts, SupportsPP
62
from .utils import (PPMissingLayer, is_pp_missing_parameter,
63
64
                    make_empty_intermediate_tensors_factory, make_layers,
                    maybe_prefix)
王敏's avatar
王敏 committed
65
from vllm import _custom_ops as ops
66
from vllm.utils import W8a8GetCacheJSON
67

68
os.environ['DPSK_FP16_QUICK'] = os.environ.get('DPSK_FP16_QUICK', '0')
wangding zeng's avatar
wangding zeng committed
69
70
71
72
73
74
75
76
77
class DeepseekV2MLP(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str,
        quant_config: Optional[QuantizationConfig] = None,
        reduce_results: bool = True,
78
        prefix: str = "",
wangding zeng's avatar
wangding zeng committed
79
80
81
82
83
    ) -> None:
        super().__init__()
        self.gate_up_proj = MergedColumnParallelLinear(
            hidden_size, [intermediate_size] * 2,
            bias=False,
84
85
            quant_config=quant_config,
            prefix=f"{prefix}.gate_up_proj")
wangding zeng's avatar
wangding zeng committed
86
87
88
89
        self.down_proj = RowParallelLinear(intermediate_size,
                                           hidden_size,
                                           bias=False,
                                           quant_config=quant_config,
90
91
                                           reduce_results=reduce_results,
                                           prefix=f"{prefix}.down_proj")
wangding zeng's avatar
wangding zeng committed
92
93
94
95
96
        if hidden_act != "silu":
            raise ValueError(f"Unsupported activation: {hidden_act}. "
                             "Only silu is supported for now.")
        self.act_fn = SiluAndMul()

97
98
99
100
101
102
103
    def forward(self, x,
                rms_weight: Optional[torch.Tensor] = None,
                residual: Optional[torch.Tensor] = None,
                update_hd: Optional[bool] = False
                ):
        if envs.USE_FUSED_RMS_QUANT:
            gate_up, new_resi, _  = self.gate_up_proj(x, rms_weight, residual, update_hd=update_hd)
104
105
106
107
108
109
            if envs.USE_FUSED_SILU_MUL_QUANT:
                x, _ = self.down_proj(gate_up, use_fused_silu_mul_quant=True)
            else:
                x = self.act_fn(gate_up)
                x, _ = self.down_proj(x)
                
110
111
112
113
114
115
            return x, new_resi
        else:
            gate_up, _ = self.gate_up_proj(x)
            x = self.act_fn(gate_up)
            x, _ = self.down_proj(x)
            return x
wangding zeng's avatar
wangding zeng committed
116
117
118
119
120
121
122
123


class DeepseekV2MoE(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        quant_config: Optional[QuantizationConfig] = None,
124
        prefix: str = "",
125
        enable_eplb: bool = False,
wangding zeng's avatar
wangding zeng committed
126
127
128
129
    ):
        super().__init__()
        self.tp_size = get_tensor_model_parallel_world_size()
        self.routed_scaling_factor = config.routed_scaling_factor
130
131
132
133
134
135

        self.ep_group = get_ep_group().device_group
        self.ep_rank = self.ep_group.rank()
        self.ep_size = self.ep_group.size()
        self.n_routed_experts: int = config.n_routed_experts
        self.n_shared_experts: int = config.n_shared_experts
136
137
138
139
140

        if config.hidden_act != "silu":
            raise ValueError(f"Unsupported activation: {config.hidden_act}. "
                             "Only silu is supported for now.")

wangding zeng's avatar
wangding zeng committed
141
        self.gate = ReplicatedLinear(config.hidden_size,
142
                                     config.n_routed_experts,
wangding zeng's avatar
wangding zeng committed
143
                                     bias=False,
144
145
                                     quant_config=None,
                                     prefix=f"{prefix}.gate")
146
147
148
149
150
151
        if config.topk_method == "noaux_tc":
            self.gate.e_score_correction_bias = nn.Parameter(
                torch.empty(config.n_routed_experts))
        else:
            self.gate.e_score_correction_bias = None

152
153
154
155
        # Load balancing settings.
        vllm_config = get_current_vllm_config()
        parallel_config = vllm_config.parallel_config
        self.enable_eplb = enable_eplb
156
        self.dpsk_fp16_quick = os.environ.get('DPSK_FP16_QUICK') == '1'
157
158
159
160
161
162
163
164
165
166
167

        self.n_redundant_experts = parallel_config.num_redundant_experts
        self.n_logical_experts = self.n_routed_experts
        self.n_physical_experts = (self.n_logical_experts +
                                   self.n_redundant_experts)
        self.n_local_physical_experts = self.n_physical_experts // self.ep_size

        self.physical_expert_start = (self.ep_rank *
                                      self.n_local_physical_experts)
        self.physical_expert_end = (self.physical_expert_start +
                                    self.n_local_physical_experts)
王敏's avatar
王敏 committed
168
        
王敏's avatar
王敏 committed
169
        self.experts = FusedMoE(
170
171
172
173
174
175
176
177
178
179
180
181
            num_experts=config.n_routed_experts,
            top_k=config.num_experts_per_tok,
            hidden_size=config.hidden_size,
            intermediate_size=config.moe_intermediate_size,
            reduce_results=False,
            renormalize=config.norm_topk_prob,
            quant_config=quant_config,
            use_grouped_topk=True,
            num_expert_group=config.n_group,
            topk_group=config.topk_group,
            prefix=f"{prefix}.experts",
            scoring_func=config.scoring_func,
王敏's avatar
王敏 committed
182
            e_score_correction_bias=self.gate.e_score_correction_bias,
183
            enable_eplb=self.enable_eplb,
zhuwenwen's avatar
zhuwenwen committed
184
            num_redundant_experts=self.n_redundant_experts,
王敏's avatar
王敏 committed
185
            routed_scaling_factor=self.routed_scaling_factor)
186

wangding zeng's avatar
wangding zeng committed
187
188
189
        if config.n_shared_experts is not None:
            intermediate_size = (config.moe_intermediate_size *
                                 config.n_shared_experts)
王敏's avatar
王敏 committed
190
            self.shared_experts = DeepseekV2MLP(
wangding zeng's avatar
wangding zeng committed
191
192
193
194
                hidden_size=config.hidden_size,
                intermediate_size=intermediate_size,
                hidden_act=config.hidden_act,
                quant_config=quant_config,
195
196
                reduce_results=self.experts.must_reduce_shared_expert_outputs(
                ),
197
                prefix=f"{prefix}.shared_experts",
wangding zeng's avatar
wangding zeng committed
198
            )
王敏's avatar
王敏 committed
199

200
        from vllm.two_batch_overlap.two_batch_overlap import tbo_all_reduce
lizhigong's avatar
lizhigong committed
201
        self.tbo_all_reduce = tbo_all_reduce
wangding zeng's avatar
wangding zeng committed
202

203
204
205
206
    def forward(self, hidden_states: torch.Tensor,
                rms_weight: Optional[torch.Tensor] = None,
                residual: Optional[torch.Tensor] = None
                ) -> torch.Tensor:
wangding zeng's avatar
wangding zeng committed
207
208
        num_tokens, hidden_dim = hidden_states.shape
        hidden_states = hidden_states.view(-1, hidden_dim)
209
        
王敏's avatar
王敏 committed
210
211
212
213
214
        if self.n_shared_experts is not None:
            if envs.USE_FUSED_RMS_QUANT:
                shared_output, new_resi = self.shared_experts(hidden_states, rms_weight, residual, update_hd=True)
            else:
                shared_output = self.shared_experts(hidden_states)
王敏's avatar
王敏 committed
215

wangding zeng's avatar
wangding zeng committed
216
217
        # router_logits: (num_tokens, n_experts)
        router_logits, _ = self.gate(hidden_states)
218

王敏's avatar
王敏 committed
219
220
221
222
223
224
225
        if envs.VLLM_USE_LIGHTOP and not self.dpsk_fp16_quick:
            final_hidden_states = self.experts(
                hidden_states=hidden_states,
                router_logits=router_logits,
                shared_output=shared_output)
        else:
            if hidden_states.dtype != torch.float16 or self.dpsk_fp16_quick:
226
227
                final_hidden_states = self.experts(
                    hidden_states=hidden_states,
王敏's avatar
王敏 committed
228
                    router_logits=router_logits) * self.routed_scaling_factor
229
            else:
王敏's avatar
王敏 committed
230
231
232
                # Fix FP16 overflow
                # See DeepseekV2DecoderLayer for more details.
                final_hidden_states = self.experts(hidden_states=hidden_states,
233
                                                router_logits=router_logits)
王敏's avatar
王敏 committed
234
        
王敏's avatar
王敏 committed
235
236
237
238
239
240
241
242
        if shared_output is not None:
            if hidden_states.dtype != torch.float16 or self.dpsk_fp16_quick:
                final_hidden_states = final_hidden_states + shared_output
            else:
                # Fix FP16 overflow
                # See DeepseekV2DecoderLayer for more details.
                final_hidden_states = final_hidden_states + shared_output \
                    * (1. / self.routed_scaling_factor)
243

王敏's avatar
王敏 committed
244
245
246
        if self.tp_size > 1:
            if envs.VLLM_ENABLE_TBO:
                final_hidden_states = self.tbo_all_reduce(final_hidden_states)
lizhigong's avatar
lizhigong committed
247
            else:
王敏's avatar
王敏 committed
248
249
250
251
252
253
254
255
                final_hidden_states = (
                    self.experts.maybe_all_reduce_tensor_model_parallel(
                        final_hidden_states))

        if envs.USE_FUSED_RMS_QUANT:
            return final_hidden_states.view(num_tokens, hidden_dim), new_resi
        else:
            return final_hidden_states.view(num_tokens, hidden_dim)
wangding zeng's avatar
wangding zeng committed
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277


def yarn_get_mscale(scale: float = 1, mscale: float = 1) -> float:
    import math
    if scale <= 1:
        return 1.0
    return 0.1 * mscale * math.log(scale) + 1.0


class DeepseekV2Attention(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        hidden_size: int,
        num_heads: int,
        qk_nope_head_dim: int,
        qk_rope_head_dim: int,
        v_head_dim: int,
        q_lora_rank: int,
        kv_lora_rank: int,
        rope_theta: float = 10000,
278
        rope_scaling: Optional[dict[str, Any]] = None,
wangding zeng's avatar
wangding zeng committed
279
280
281
        max_position_embeddings: int = 8192,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
282
        prefix: str = "",
wangding zeng's avatar
wangding zeng committed
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
    ) -> None:
        super().__init__()
        self.hidden_size = hidden_size
        self.qk_nope_head_dim = qk_nope_head_dim
        self.qk_rope_head_dim = qk_rope_head_dim
        self.qk_head_dim = qk_nope_head_dim + qk_rope_head_dim
        self.v_head_dim = v_head_dim
        self.q_lora_rank = q_lora_rank
        self.kv_lora_rank = kv_lora_rank
        self.num_heads = num_heads
        tp_size = get_tensor_model_parallel_world_size()
        assert num_heads % tp_size == 0
        self.num_local_heads = num_heads // tp_size
        self.scaling = self.qk_head_dim**-0.5
        self.rope_theta = rope_theta
        self.max_position_embeddings = max_position_embeddings

        if self.q_lora_rank is not None:
            self.q_a_proj = ReplicatedLinear(self.hidden_size,
                                             self.q_lora_rank,
                                             bias=False,
304
305
                                             quant_config=quant_config,
                                             prefix=f"{prefix}.q_a_proj")
wangding zeng's avatar
wangding zeng committed
306
307
308
309
310
311
            self.q_a_layernorm = RMSNorm(self.q_lora_rank,
                                         eps=config.rms_norm_eps)
            self.q_b_proj = ColumnParallelLinear(q_lora_rank,
                                                 self.num_heads *
                                                 self.qk_head_dim,
                                                 bias=False,
312
313
                                                 quant_config=quant_config,
                                                 prefix=f"{prefix}.q_b_proj")
wangding zeng's avatar
wangding zeng committed
314
315
316
317
318
        else:
            self.q_proj = ColumnParallelLinear(self.hidden_size,
                                               self.num_heads *
                                               self.qk_head_dim,
                                               bias=False,
319
320
                                               quant_config=quant_config,
                                               prefix=f"{prefix}.q_proj")
wangding zeng's avatar
wangding zeng committed
321

322
323
324
325
326
327
        self.kv_a_proj_with_mqa = ReplicatedLinear(
            self.hidden_size,
            self.kv_lora_rank + self.qk_rope_head_dim,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.kv_a_proj_with_mqa")
wangding zeng's avatar
wangding zeng committed
328
329
330
331
332
333
        self.kv_a_layernorm = RMSNorm(self.kv_lora_rank,
                                      eps=config.rms_norm_eps)
        self.kv_b_proj = ColumnParallelLinear(
            self.kv_lora_rank,
            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
            bias=False,
334
335
            quant_config=quant_config,
            prefix=f"{prefix}.kv_b_proj")
wangding zeng's avatar
wangding zeng committed
336
337
338
339
        # O projection.
        self.o_proj = RowParallelLinear(self.num_heads * self.v_head_dim,
                                        self.hidden_size,
                                        bias=False,
340
341
                                        quant_config=quant_config,
                                        prefix=f"{prefix}.o_proj")
342
343
        if rope_scaling:
            rope_scaling["rope_type"] = 'deepseek_yarn'
344

wangding zeng's avatar
wangding zeng committed
345
346
347
348
349
350
351
352
353
354
355
356
357
358
        self.rotary_emb = get_rope(qk_rope_head_dim,
                                   rotary_dim=qk_rope_head_dim,
                                   max_position=max_position_embeddings,
                                   base=rope_theta,
                                   rope_scaling=rope_scaling,
                                   is_neox_style=False)

        if rope_scaling:
            mscale_all_dim = rope_scaling.get("mscale_all_dim", False)
            scaling_factor = rope_scaling["factor"]
            mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
            self.scaling = self.scaling * mscale * mscale

        self.attn = Attention(self.num_local_heads,
359
                              self.qk_head_dim,
wangding zeng's avatar
wangding zeng committed
360
361
362
                              self.scaling,
                              num_kv_heads=self.num_local_heads,
                              cache_config=cache_config,
363
364
                              quant_config=quant_config,
                              prefix=f"{prefix}.attn")
wangding zeng's avatar
wangding zeng committed
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        if self.q_lora_rank is not None:
            q = self.q_a_proj(hidden_states)[0]
            q = self.q_a_layernorm(q)
            q = self.q_b_proj(q)[0].view(-1, self.num_local_heads,
                                         self.qk_head_dim)
        else:
            q = self.q_proj(hidden_states)[0].view(-1, self.num_local_heads,
                                                   self.qk_head_dim)
        q_nope, q_pe = q.split([self.qk_nope_head_dim, self.qk_rope_head_dim],
                               dim=-1)
        latent_cache = self.kv_a_proj_with_mqa(hidden_states)[0]
        kv_a, _ = latent_cache.split(
            [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
        latent_cache = latent_cache.unsqueeze(1)
        kv_a = self.kv_a_layernorm(kv_a.contiguous())
        kv = self.kv_b_proj(kv_a)[0]
        kv = kv.view(-1, self.num_local_heads,
                     self.qk_nope_head_dim + self.v_head_dim)
        k_nope, v = kv.split([self.qk_nope_head_dim, self.v_head_dim], dim=-1)
        k_pe = latent_cache[:, :, self.kv_lora_rank:]
391

wangding zeng's avatar
wangding zeng committed
392
        q_pe, k_pe = self.rotary_emb(positions, q_pe, k_pe)
393

wangding zeng's avatar
wangding zeng committed
394
395
396
397
        q[..., self.qk_nope_head_dim:] = q_pe
        k = torch.empty_like(q)
        k[..., :self.qk_nope_head_dim] = k_nope
        k[..., self.qk_nope_head_dim:] = k_pe
398
399
400
401
        # padding value to qk_head_dim for alignment
        v = torch.nn.functional.pad(
            v, [0, self.qk_head_dim - self.v_head_dim],
            value=0).view(-1, self.num_local_heads * self.qk_head_dim)
402
        attn_output = self.attn(q, k, v)
wangding zeng's avatar
wangding zeng committed
403
        attn_output = attn_output.view(
404
405
            -1, self.num_local_heads,
            self.qk_head_dim)[..., :self.v_head_dim].reshape(
wangding zeng's avatar
wangding zeng committed
406
407
408
409
410
                -1, self.num_local_heads * self.v_head_dim)
        output, _ = self.o_proj(attn_output)
        return output


411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
class DeepseekV2MLAAttention(nn.Module):
    """
    Main reference: DeepseekV2 paper, and FlashInfer Implementation
    (https://arxiv.org/abs/2405.04434 and https://github.com/flashinfer-ai/flashinfer/pull/551).
    
    For more info see MLACommonImpl in: vllm/attention/backends/mla/utils.py
    """

    def __init__(
        self,
        config: PretrainedConfig,
        hidden_size: int,
        num_heads: int,
        qk_nope_head_dim: int,
        qk_rope_head_dim: int,
        v_head_dim: int,
        q_lora_rank: Optional[int],
        kv_lora_rank: int,
        rope_theta: float = 10000,
430
        rope_scaling: Optional[dict[str, Any]] = None,
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
        max_position_embeddings: int = 8192,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.hidden_size = hidden_size
        self.qk_nope_head_dim = qk_nope_head_dim
        self.qk_rope_head_dim = qk_rope_head_dim
        self.qk_head_dim = qk_nope_head_dim + qk_rope_head_dim
        self.v_head_dim = v_head_dim

        self.q_lora_rank = q_lora_rank
        self.kv_lora_rank = kv_lora_rank

        self.num_heads = num_heads
        tp_size = get_tensor_model_parallel_world_size()
        assert num_heads % tp_size == 0
        self.num_local_heads = num_heads // tp_size

        self.scaling = self.qk_head_dim**-0.5
        self.rope_theta = rope_theta
        self.max_position_embeddings = max_position_embeddings

        if self.q_lora_rank is not None:
456
457
            if envs.USE_FUSED_RMS_QUANT:
                self.q_a_proj = ReplicatedLinear(self.hidden_size,
458
459
460
                                             self.q_lora_rank,
                                             bias=False,
                                             quant_config=quant_config,
461
                                             eps=config.rms_norm_eps,
462
                                             prefix=f"{prefix}.q_a_proj")
463
                self.q_b_proj = ColumnParallelLinear(q_lora_rank,
464
465
466
467
                                                 self.num_heads *
                                                 self.qk_head_dim,
                                                 bias=False,
                                                 quant_config=quant_config,
468
                                                 eps=config.rms_norm_eps,
469
                                                 prefix=f"{prefix}.q_b_proj")
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
            else:
                self.q_a_proj = ReplicatedLinear(self.hidden_size,
                                             self.q_lora_rank,
                                             bias=False,
                                             quant_config=quant_config,
                                             prefix=f"{prefix}.q_a_proj")
                self.q_b_proj = ColumnParallelLinear(q_lora_rank,
                                                 self.num_heads *
                                                 self.qk_head_dim,
                                                 bias=False,
                                                 quant_config=quant_config,
                                                 prefix=f"{prefix}.q_b_proj")
                
            self.q_a_layernorm = RMSNorm(self.q_lora_rank,
                                         eps=config.rms_norm_eps)

486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
        else:
            self.q_proj = ColumnParallelLinear(self.hidden_size,
                                               self.num_heads *
                                               self.qk_head_dim,
                                               bias=False,
                                               quant_config=quant_config,
                                               prefix=f"{prefix}.q_proj")

        self.kv_a_proj_with_mqa = ReplicatedLinear(
            self.hidden_size,
            self.kv_lora_rank + self.qk_rope_head_dim,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.kv_a_proj_with_mqa")
        self.kv_a_layernorm = RMSNorm(self.kv_lora_rank,
                                      eps=config.rms_norm_eps)
        self.kv_b_proj = ColumnParallelLinear(
            self.kv_lora_rank,
            self.num_heads * (self.qk_nope_head_dim + self.v_head_dim),
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.kv_b_proj")
        self.o_proj = RowParallelLinear(self.num_heads * self.v_head_dim,
                                        self.hidden_size,
                                        bias=False,
                                        quant_config=quant_config,
                                        prefix=f"{prefix}.o_proj")

514
515
        if rope_scaling:
            rope_scaling["rope_type"] = 'deepseek_yarn'
516
517
518
519
520
521
522
523
524
525
526
527
        self.rotary_emb = get_rope(qk_rope_head_dim,
                                   rotary_dim=qk_rope_head_dim,
                                   max_position=max_position_embeddings,
                                   base=rope_theta,
                                   rope_scaling=rope_scaling,
                                   is_neox_style=False)
        if rope_scaling:
            mscale_all_dim = rope_scaling.get("mscale_all_dim", False)
            scaling_factor = rope_scaling["factor"]
            mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
            self.scaling = self.scaling * mscale * mscale

528
529
530
531
532
533
        # In the MLA backend, kv_cache includes both k_c and
        # pe (i.e. decoupled position embeddings). In particular,
        # the concat_and_cache_mla op requires
        #     k_c.size(1) + k_pe.size(1) == kv_cache.size(2)
        # i.e.
        #     kv_lora_rank + qk_rope_head_dim == head_size
534
535
        self.mla_attn = Attention(
            num_heads=self.num_local_heads,
536
            head_size=self.kv_lora_rank + self.qk_rope_head_dim,
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
            scale=self.scaling,
            num_kv_heads=1,
            cache_config=cache_config,
            quant_config=quant_config,
            prefix=f"{prefix}.attn",
            use_mla=True,
            # MLA Args
            q_lora_rank=self.q_lora_rank,
            kv_lora_rank=self.kv_lora_rank,
            qk_nope_head_dim=self.qk_nope_head_dim,
            qk_rope_head_dim=self.qk_rope_head_dim,
            qk_head_dim=self.qk_head_dim,
            v_head_dim=self.v_head_dim,
            kv_b_proj=self.kv_b_proj,
        )

        self.prefix = prefix
        self.debug_layer_idx = int(self.prefix.split(".")[-2])

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
560
561
        rms_weight: Optional[torch.Tensor] = None,
        residual: Optional[torch.Tensor] = None
562
    ) -> torch.Tensor:
563
564
565
566
567
568
569
570
571
572
        if envs.USE_FUSED_RMS_QUANT and rms_weight is not None:
            if self.q_lora_rank is not None:
                q_c, new_residual, _, input_quant_args = self.q_a_proj(hidden_states, rms_weight=rms_weight, residual=residual, update_hd=False)
                q, _, _ = self.q_b_proj(q_c, rms_weight=self.q_a_layernorm.weight.data, residual=None, update_hd=False)
                
            else:
                q = self.q_proj(hidden_states)[0]
            kv_c, k_pe = self.kv_a_proj_with_mqa(hidden_states, quant_args=input_quant_args, update_hd=False)[0].split(
                                [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
            
573
574
575
576
            if envs.VLLM_USE_LIGHTOP:
                kv_c_normed = self.kv_a_layernorm.forward_cuda_opt(kv_c)
            else:
                kv_c_normed = self.kv_a_layernorm(kv_c.contiguous())
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591

            q = q.view(-1, self.num_local_heads, self.qk_head_dim)
            # Add head dim of 1 to k_pe
            k_pe = k_pe.unsqueeze(1)

            q[..., self.qk_nope_head_dim:], k_pe = self.rotary_emb(
                positions, q[..., self.qk_nope_head_dim:], k_pe)

            attn_out = self.mla_attn(
                q,
                kv_c_normed,
                k_pe,
                output_shape=(hidden_states.shape[0],
                            self.num_local_heads * self.v_head_dim))
            return self.o_proj(attn_out)[0], new_residual
592
        else:
593
594
595
596
597
598
599
600
601
            if self.q_lora_rank is not None:
                q_c = self.q_a_proj(hidden_states)[0]
                q_c = self.q_a_layernorm(q_c)
                q = self.q_b_proj(q_c)[0]
            else:
                q = self.q_proj(hidden_states)[0]
            kv_c, k_pe = self.kv_a_proj_with_mqa(hidden_states)[0].split(
                [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)
            kv_c_normed = self.kv_a_layernorm(kv_c.contiguous())
602

603
604
605
            q = q.view(-1, self.num_local_heads, self.qk_head_dim)
            # Add head dim of 1 to k_pe
            k_pe = k_pe.unsqueeze(1)
606

607
608
            q[..., self.qk_nope_head_dim:], k_pe = self.rotary_emb(
                positions, q[..., self.qk_nope_head_dim:], k_pe)
609

610
611
612
613
614
615
616
            attn_out = self.mla_attn(
                q,
                kv_c_normed,
                k_pe,
                output_shape=(hidden_states.shape[0],
                            self.num_local_heads * self.v_head_dim))
            return self.o_proj(attn_out)[0]
617
618


wangding zeng's avatar
wangding zeng committed
619
620
621
622
623
class DeepseekV2DecoderLayer(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
624
        prefix: str,
625
        model_config: ModelConfig,
wangding zeng's avatar
wangding zeng committed
626
627
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
628
        enable_eplb: bool = False,
wangding zeng's avatar
wangding zeng committed
629
630
631
632
633
634
635
    ) -> None:
        super().__init__()
        self.hidden_size = config.hidden_size
        rope_theta = getattr(config, "rope_theta", 10000)
        rope_scaling = getattr(config, "rope_scaling", None)
        max_position_embeddings = getattr(config, "max_position_embeddings",
                                          8192)
636
637
638
        # DecoderLayers are created with `make_layers` which passes the prefix
        # with the layer's index.
        layer_idx = int(prefix.split(sep='.')[-1])
639
        self.layer_idx = layer_idx
640
641
642
643
644
        if model_config.use_mla:
            attn_cls = DeepseekV2MLAAttention
        else:
            attn_cls = DeepseekV2Attention
        self.self_attn = attn_cls(
wangding zeng's avatar
wangding zeng committed
645
646
647
648
649
650
651
652
653
654
655
656
657
658
            config=config,
            hidden_size=self.hidden_size,
            num_heads=config.num_attention_heads,
            qk_nope_head_dim=config.qk_nope_head_dim,
            qk_rope_head_dim=config.qk_rope_head_dim,
            v_head_dim=config.v_head_dim,
            q_lora_rank=config.q_lora_rank
            if hasattr(config, "q_lora_rank") else None,
            kv_lora_rank=config.kv_lora_rank,
            rope_theta=rope_theta,
            rope_scaling=rope_scaling,
            max_position_embeddings=max_position_embeddings,
            cache_config=cache_config,
            quant_config=quant_config,
659
            prefix=f"{prefix}.self_attn",
wangding zeng's avatar
wangding zeng committed
660
        )
661
        self.dpsk_fp16_quick = os.environ.get('DPSK_FP16_QUICK') == '1'
wangding zeng's avatar
wangding zeng committed
662
663
664
        if (config.n_routed_experts is not None
                and layer_idx >= config.first_k_dense_replace
                and layer_idx % config.moe_layer_freq == 0):
665
666
667
668
            self.mlp = DeepseekV2MoE(
                config=config,
                quant_config=quant_config,
                prefix=f"{prefix}.mlp",
669
                enable_eplb=enable_eplb,
670
            )
wangding zeng's avatar
wangding zeng committed
671
672
673
674
675
676
        else:
            self.mlp = DeepseekV2MLP(
                hidden_size=config.hidden_size,
                intermediate_size=config.intermediate_size,
                hidden_act=config.hidden_act,
                quant_config=quant_config,
677
                prefix=f"{prefix}.mlp",
wangding zeng's avatar
wangding zeng committed
678
679
680
681
682
            )
        self.input_layernorm = RMSNorm(config.hidden_size,
                                       eps=config.rms_norm_eps)
        self.post_attention_layernorm = RMSNorm(config.hidden_size,
                                                eps=config.rms_norm_eps)
683
        self.routed_scaling_factor = config.routed_scaling_factor
wangding zeng's avatar
wangding zeng committed
684
685
686
687
688
689
690

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor],
    ) -> torch.Tensor:
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
        if envs.USE_FUSED_RMS_QUANT:
            # Fix residual FP16 overflow
            residual_fix_overflow = False
            
            assert self.input_layernorm.has_weight is True
            if residual is None:
                residual = hidden_states
                hidden_states, _ = self.self_attn(
                    positions = positions,
                    hidden_states = hidden_states,
                    rms_weight = self.input_layernorm.weight.data,
                    residual = None
                )
                residual_fix_overflow = True
            else:
                hidden_states, new_residual = self.self_attn(
                    positions = positions,
                    hidden_states = hidden_states,
                    rms_weight = self.input_layernorm.weight.data,
                    residual = residual
                )
                residual = new_residual
                
            if hidden_states.dtype == torch.float16 and not self.dpsk_fp16_quick:
                # rmsnorm, and rmsnorm result would not affect by scale.
                hidden_states *= 1. / self.routed_scaling_factor
                if self.layer_idx == 0 or residual_fix_overflow:
                    # The residual is shared by all layers, we only scale it on
                    # first layer.
                    residual *= 1. / self.routed_scaling_factor

            hidden_states, new_resi = self.mlp(hidden_states, self.post_attention_layernorm.weight.data, residual)

            if isinstance(self.mlp,
                        DeepseekV2MLP) and hidden_states.dtype == torch.float16 and not self.dpsk_fp16_quick:
                # Fix FP16 overflow
                # Scaling the DeepseekV2MLP output, it is the input of
                # input_layernorm of next decoder layer.
                # The scaling of DeepseekV2MOE output would be done in the forward
                # of DeepseekV2MOE
                hidden_states *= 1. / self.routed_scaling_factor
            return hidden_states, new_resi

wangding zeng's avatar
wangding zeng committed
734
        else:
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
            # Self Attention
            # Fix residual FP16 overflow
            residual_fix_overflow = False
            if residual is None:
                residual = hidden_states
                hidden_states = self.input_layernorm(hidden_states)
                residual_fix_overflow = True
            else:
                hidden_states, residual = self.input_layernorm(
                    hidden_states, residual)
            hidden_states = self.self_attn(
                positions=positions,
                hidden_states=hidden_states,
            )

            if hidden_states.dtype == torch.float16 and not self.dpsk_fp16_quick:
                # Fix FP16 overflow
                # We scale both hidden_states and residual before
                # rmsnorm, and rmsnorm result would not affect by scale.
                hidden_states *= 1. / self.routed_scaling_factor
                if self.layer_idx == 0 or residual_fix_overflow:
                    # The residual is shared by all layers, we only scale it on
                    # first layer.
                    residual *= 1. / self.routed_scaling_factor

            # Fully Connected
            hidden_states, residual = self.post_attention_layernorm(
wangding zeng's avatar
wangding zeng committed
762
                hidden_states, residual)
763
            hidden_states = self.mlp(hidden_states)
wangding zeng's avatar
wangding zeng committed
764

765
766
767
768
769
770
771
772
            if isinstance(self.mlp,
                        DeepseekV2MLP) and hidden_states.dtype == torch.float16 and not self.dpsk_fp16_quick:
                # Fix FP16 overflow
                # Scaling the DeepseekV2MLP output, it is the input of
                # input_layernorm of next decoder layer.
                # The scaling of DeepseekV2MOE output would be done in the forward
                # of DeepseekV2MOE
                hidden_states *= 1. / self.routed_scaling_factor
773

774
            return hidden_states, residual
wangding zeng's avatar
wangding zeng committed
775
776


777
@support_torch_compile
wangding zeng's avatar
wangding zeng committed
778
779
780
781
class DeepseekV2Model(nn.Module):

    fall_back_to_pt_during_load = False

782
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
wangding zeng's avatar
wangding zeng committed
783
        super().__init__()
784
785

        config = vllm_config.model_config.hf_config
786
        model_config = vllm_config.model_config
787
788
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config
789
        enable_eplb = vllm_config.parallel_config.enable_eplb
790
        self.config = config
791

wangding zeng's avatar
wangding zeng committed
792
793
        self.vocab_size = config.vocab_size

794
795
796
797
        if get_pp_group().is_first_rank:
            self.embed_tokens = VocabParallelEmbedding(
                config.vocab_size,
                config.hidden_size,
798
799
                quant_config=quant_config,
                prefix=f"{prefix}.embed_tokens")
800
801
802
803
804
805
806
807
        else:
            self.embed_tokens = PPMissingLayer()

        self.start_layer, self.end_layer, self.layers = make_layers(
            config.num_hidden_layers,
            lambda prefix: DeepseekV2DecoderLayer(
                config,
                prefix,
808
                model_config=model_config,
809
810
                cache_config=cache_config,
                quant_config=quant_config,
811
                enable_eplb=enable_eplb,
812
813
814
815
816
817
818
            ),
            prefix=f"{prefix}.layers")

        if get_pp_group().is_last_rank:
            self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        else:
            self.norm = PPMissingLayer()
819
820
821
        self.make_empty_intermediate_tensors = (
            make_empty_intermediate_tensors_factory(
                ["hidden_states", "residual"], config.hidden_size))
wangding zeng's avatar
wangding zeng committed
822

823
824
825
    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.embed_tokens(input_ids)

wangding zeng's avatar
wangding zeng committed
826
827
828
829
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
830
        intermediate_tensors: Optional[IntermediateTensors],
831
        inputs_embeds: Optional[torch.Tensor] = None,
832
    ) -> Union[torch.Tensor, IntermediateTensors]:
833
        if get_pp_group().is_first_rank:
834
835
836
837
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
                hidden_states = self.get_input_embeddings(input_ids)
838
839
840
841
842
843
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]

844
        for layer in self.layers[self.start_layer:self.end_layer]:
845
            hidden_states, residual = layer(positions, hidden_states, residual)
846
847
848
849
850
851
852

        if not get_pp_group().is_last_rank:
            return IntermediateTensors({
                "hidden_states": hidden_states,
                "residual": residual
            })

wangding zeng's avatar
wangding zeng committed
853
854
855
856
        hidden_states, _ = self.norm(hidden_states, residual)
        return hidden_states


857
class DeepseekV2ForCausalLM(nn.Module, SupportsPP, MixtureOfExperts):
wangding zeng's avatar
wangding zeng committed
858

859
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
wangding zeng's avatar
wangding zeng committed
860
        super().__init__()
861
862
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
863
864
865
866
867
868
869

        self.quant_method = None
        if quant_config is not None:
            self.quant_method = quant_config.get_name()
            os.environ['LLAMA_NN'] = '0'
            os.environ['LM_NN'] = '0'

870
        self.use_w4a16_moe_sz = os.environ.get('AWQ_MOE_SZ') == '1'
wangding zeng's avatar
wangding zeng committed
871
872
        self.config = config
        self.quant_config = quant_config
王敏's avatar
王敏 committed
873

874
        self.model = DeepseekV2Model(vllm_config=vllm_config,
875
                                     prefix=maybe_prefix(prefix, "model"))
876
877
878
879
880
881
        if get_pp_group().is_last_rank:
            self.lm_head = ParallelLMHead(config.vocab_size,
                                          config.hidden_size,
                                          quant_config=quant_config)
        else:
            self.lm_head = PPMissingLayer()
wangding zeng's avatar
wangding zeng committed
882
        self.logits_processor = LogitsProcessor(config.vocab_size)
883
884
        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors)
885
886
887
888
889
890
891
892
        self.expert_weights = []

        # Set MoE hyperparameters
        self.num_moe_layers = (config.num_hidden_layers -
                               config.first_k_dense_replace)
        self.num_expert_groups = config.n_group

        self.moe_layers: list[FusedMoE] = []
893
        example_moe = None
894
        for layer in self.model.layers:
895
896
897
            if isinstance(layer, PPMissingLayer):
                continue

898
899
            assert isinstance(layer, DeepseekV2DecoderLayer)
            if isinstance(layer.mlp, DeepseekV2MoE):
900
                example_moe = layer.mlp
901
902
903
904
905
906
907
908
909
                self.moe_layers.append(layer.mlp.experts)

        # Pick last one layer since the first ones may be dense layers.
        self.num_logical_experts = example_moe.n_logical_experts
        self.num_physical_experts = example_moe.n_physical_experts
        self.num_local_physical_experts = example_moe.n_local_physical_experts
        self.num_routed_experts = example_moe.n_routed_experts
        self.num_shared_experts = example_moe.n_shared_experts
        self.num_redundant_experts = example_moe.n_redundant_experts
zhuwenwen's avatar
zhuwenwen committed
910
        
王敏's avatar
王敏 committed
911
        self.use_llama_nn = os.environ.get('LLAMA_NN') == '1'
912
        self.use_awq_pad = os.environ.get('AWQ_PAD') == '1'
913
914
        self.tritonsingleton= W8a8GetCacheJSON() 
        self.tritonsingleton.topk = config.num_experts_per_tok
王敏's avatar
王敏 committed
915
        self.tritonsingleton.quant_method=self.quant_method
王敏's avatar
王敏 committed
916

917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
    def set_eplb_state(
        self,
        expert_load_view: torch.Tensor,
        logical_to_physical_map: torch.Tensor,
        logical_replica_count: torch.Tensor,
    ) -> None:
        for layer_idx, layer in enumerate(self.moe_layers):
            # Register the expert weights.
            self.expert_weights.append(layer.get_expert_weights())
            layer.set_eplb_state(
                moe_layer_idx=layer_idx,
                expert_load_view=expert_load_view,
                logical_to_physical_map=logical_to_physical_map,
                logical_replica_count=logical_replica_count,
            )
932

933
934
935
    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.model.get_input_embeddings(input_ids)

wangding zeng's avatar
wangding zeng committed
936
937
938
939
    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
940
        intermediate_tensors: Optional[IntermediateTensors] = None,
941
        inputs_embeds: Optional[torch.Tensor] = None,
942
    ) -> Union[torch.Tensor, IntermediateTensors]:
943
        hidden_states = self.model(input_ids, positions, intermediate_tensors,
944
                                   inputs_embeds)
wangding zeng's avatar
wangding zeng committed
945
946
        return hidden_states

947
948
949
950
951
    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[torch.Tensor]:
952
        logits = self.logits_processor(self.lm_head, hidden_states,
wangding zeng's avatar
wangding zeng committed
953
954
955
                                       sampling_metadata)
        return logits

956
957
958
959
960
961
962
963
964
965
966
967
968
    def make_empty_intermediate_tensors(
            self, batch_size: int, dtype: torch.dtype,
            device: torch.device) -> IntermediateTensors:
        return IntermediateTensors({
            "hidden_states":
            torch.zeros((batch_size, self.config.hidden_size),
                        dtype=dtype,
                        device=device),
            "residual":
            torch.zeros((batch_size, self.config.hidden_size),
                        dtype=dtype,
                        device=device),
        })
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
        
    def restore_qzeros_tensor(self, qzeros, qscales):

        low_bits = qzeros & 0x0F
        high_bits = qzeros >> 4
        
        zeors_tensor = torch.stack([low_bits, high_bits], dim=2).view(qzeros.shape[0], -1 , qzeros.shape[-1])
        zeors_int16 = zeors_tensor.to(torch.int16)
        assert zeors_int16.shape == qscales.shape

        uint16_tensor1 = zeors_int16.view(torch.uint16)
        uint16_tensor2 = qscales.view(torch.uint16)
        
        uint32_tensor1 = uint16_tensor1.to(torch.int32) << 16
        uint32_tensor2 = uint16_tensor2.to(torch.int32)
        
        result_tensor = uint32_tensor1 + uint32_tensor2
        result_tensor =result_tensor.view(torch.uint32)
        result_tensor = result_tensor.transpose(1, 2).contiguous()
        return result_tensor
989

990
991
    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
wangding zeng's avatar
wangding zeng committed
992
993
994
995
996
997
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("gate_up_proj", "gate_proj", 0),
            ("gate_up_proj", "up_proj", 1),
        ]

998
999
        # Params for weights, fp8 weight scales, fp8 activation scales
        # (param_name, weight_name, expert_id, shard_id)
王敏's avatar
王敏 committed
1000
1001
1002
1003
        expert_params_mapping = FusedMoE.make_expert_params_mapping(
            ckpt_gate_proj_name="gate_proj",
            ckpt_down_proj_name="down_proj",
            ckpt_up_proj_name="up_proj",
1004
1005
            num_experts=self.config.n_routed_experts,
            num_redundant_experts=self.num_redundant_experts)
1006

wangding zeng's avatar
wangding zeng committed
1007
        params_dict = dict(self.named_parameters())
1008
        loaded_params: set[str] = set()
wangding zeng's avatar
wangding zeng committed
1009
1010
1011
        for name, loaded_weight in weights:
            if "rotary_emb.inv_freq" in name:
                continue
1012

1013
1014
1015
            spec_layer = get_spec_layer_idx_from_weight_name(self.config, name)
            if spec_layer is not None:
                continue  # skip spec decode layers for main model
1016

wangding zeng's avatar
wangding zeng committed
1017
            for (param_name, weight_name, shard_id) in stacked_params_mapping:
1018
                # Skip non-stacked layers and experts (experts handled below).
wangding zeng's avatar
wangding zeng committed
1019
1020
                if weight_name not in name:
                    continue
1021
1022
1023
1024
1025
1026
1027
1028
                # We have mlp.experts[0].gate_proj in the checkpoint.
                # Since we handle the experts below in expert_params_mapping,
                # we need to skip here BEFORE we update the name, otherwise
                # name will be updated to mlp.experts[0].gate_up_proj, which
                # will then be updated below in expert_params_mapping
                # for mlp.experts[0].gate_gate_up_proj, which breaks load.
                if (("mlp.experts." in name) and name not in params_dict):
                    continue
wangding zeng's avatar
wangding zeng committed
1029
                name = name.replace(weight_name, param_name)
王敏's avatar
王敏 committed
1030

wangding zeng's avatar
wangding zeng committed
1031
1032
1033
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue
1034
1035
1036
1037

                if is_pp_missing_parameter(name, self):
                    continue

wangding zeng's avatar
wangding zeng committed
1038
1039
1040
1041
1042
                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
1043
                is_expert_weight = False
1044
1045
1046
1047
                for mapping in expert_params_mapping:
                    param_name, weight_name, expert_id, shard_id = mapping
                    if weight_name not in name:
                        continue
1048

1049
1050
1051
1052
1053
1054
1055
1056
1057
                    # Anyway, this is an expert weight and should not be
                    # attempted to load as other weights later
                    is_expert_weight = True

                    # Do not modify `name` since the loop may continue here
                    # Instead, create a new variable
                    name_mapped = name.replace(weight_name, param_name)

                    if is_pp_missing_parameter(name_mapped, self):
1058
1059
                        continue

1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
                    param = params_dict[name_mapped]
                    # We should ask the weight loader to return success or not
                    # here since otherwise we may skip experts with other
                    # available replicas.
                    weight_loader = typing.cast(Callable[..., bool],
                                                param.weight_loader)
                    success = weight_loader(param,
                                            loaded_weight,
                                            name_mapped,
                                            shard_id=shard_id,
                                            expert_id=expert_id,
                                            return_success=True)
                    if success:
1073
                        name = name_mapped
1074
                        break
1075
                else:
1076
1077
1078
1079
1080
                    if is_expert_weight:
                        # We've checked that this is an expert weight
                        # However it's not mapped locally to this rank
                        # So we simply skip it
                        continue
王敏's avatar
王敏 committed
1081
                    
1082
1083
1084
1085
                    # Skip loading extra bias for GPTQ models.
                    if name.endswith(".bias") and name not in params_dict:
                        continue

1086
1087
1088
1089
1090
                    # Remapping the name of FP8 kv-scale.
                    name = maybe_remap_kv_scale_name(name, params_dict)
                    if name is None:
                        continue

1091
1092
1093
                    if is_pp_missing_parameter(name, self):
                        continue

zhuwenwen's avatar
zhuwenwen committed
1094
1095
1096
1097
                    try:
                        param = params_dict[name]
                    except Exception as e:
                        continue
1098
1099
1100
                    weight_loader = getattr(param, "weight_loader",
                                            default_weight_loader)
                    weight_loader(param, loaded_weight)
1101
            loaded_params.add(name)
王敏's avatar
王敏 committed
1102
            
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
        if self.use_llama_nn and self.quant_method is None:
            lay_key_words = [
                "self_attn.q_proj.weight",
                "self_attn.q_a_proj.weight",
                "self_attn.q_b_proj.weight",
                "self_attn.kv_a_proj_with_mqa.weight",
                "self_attn.kv_b_proj.weight",
                "self_attn.o_proj.weight",
                "mlp.gate_up_proj.weight",
                "mlp.down_proj.weight",
                "mlp.gate.weight",
                "shared_experts.gate_up_proj.weight",
                "shared_experts.down_proj.weight",
                "lm_head.weight",
            ]

            combined_words = "|".join(lay_key_words)
            
            for layername in loaded_params:
                weight = params_dict[layername]
                matches = re.findall(combined_words, layername)
                if matches:
                    _weight = torch.zeros_like(weight.data)
                    ori_shape =_weight.shape
                    
                    ops.trans_w16_gemm(_weight, weight.data, _weight.shape[0], _weight.shape[1])
                    weight.data.copy_(_weight)
                    
                    weight.data=weight.data.reshape(ori_shape[1],-1)
zhuwenwen's avatar
zhuwenwen committed
1132
            
1133
        return loaded_params
1134
1135
1136
1137


class DeepseekV3ForCausalLM(DeepseekV2ForCausalLM):
    pass
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149


def get_spec_layer_idx_from_weight_name(config: PretrainedConfig,
                                        weight_name: str) -> Optional[int]:
    if hasattr(config,
               "num_nextn_predict_layers") and (config.num_nextn_predict_layers
                                                > 0):
        layer_idx = config.num_hidden_layers
        for i in range(config.num_nextn_predict_layers):
            if weight_name.startswith(f"model.layers.{layer_idx+i}."):
                return layer_idx + i
    return None